Use BiOS from Cursor

AI code editor

Works today

Cursor supports MCP servers per project or globally, and its agent picks up new tools without extra setup. One small JSON file connects it to BiOS.

Before you start

You need a BiOS API key. In your dashboard, open API Keys, click Create API Key, pick a permission preset, and copy the sk_live_ key. It is shown only once.

Node.js 18 or newer must be installed; the server runs with npx, so there is nothing else to install.

Setup

1
Create the MCP config
For one project, create .cursor/mcp.json in the project root. For every project, create ~/.cursor/mcp.json in your home directory instead. The content is the same:
.cursor/mcp.json
```
{
  "mcpServers": {
    "usfbios": {
      "command": "npx",
      "args": ["-y", "@usfbios/mcp-server"],
      "env": {
        "USFBIOS_API_KEY": "sk_live_your_api_key"
      }
    }
  }
}
```
2
Enable the server
Open Cursor Settings, then the MCP section. The usfbios server appears in the list; make sure its toggle is on. A green dot means Cursor started it and found the tools.
3
Use it in Agent mode
Open the chat panel in Agent mode and ask for a fine-tuning task. Cursor shows each tool call for approval before it runs.
text
```
> Preview dataset ds_abc123, then start an SFT job on
  Llama 3.1 8B with LoRA rank 16 and 3 epochs.
```

Check that it works

Cursor Settings, MCP shows usfbios with a green status dot and a tool count.
In an Agent chat, ask "What GPUs does BiOS offer and at what price?" and confirm it calls get_gpu_pricing.

Try these prompts

“Search for Llama models under 13B parameters and recommend a GPU for LoRA fine-tuning”

“Upload ~/data/support-qa.jsonl and start an SFT job on Llama 3.1 8B with LoRA”

“What is the status of my running training jobs? Show the loss curve for the latest one”

“Deploy my finished fine-tune and give me the endpoint URL”

Your assistant has 38 BiOS tools in total. The tool reference documents each one.

Troubleshooting

The status dot stays yellow or red

Click the server row to see its logs. The usual causes are Node.js missing (install Node 18+) or invalid JSON in mcp.json.

The agent answers from memory instead of calling tools

Make sure you are in Agent mode, not the plain chat. You can also name the tool directly: "Use the usfbios tools to list my datasets".

The server never starts, or logs show "npx: command not found"

The MCP server runs on Node.js. Install Node 18 or newer from nodejs.org, restart your client, and try again. Run npx -y @usfbios/mcp-server in a terminal to confirm it starts (it waits for input; press Ctrl+C to exit).

Tools appear but every call fails with an authentication error

Check that USFBIOS_API_KEY is set to a key that starts with sk_live_ and has not expired. Keys are shown only once at creation, so if in doubt create a fresh one under Dashboard, then API Keys.

A tool call fails with a permission error

Your key may be scoped too narrowly. Ask the assistant to call introspect_api_key to see exactly which scopes the key has, then create a key with the preset you need (Read Only, Training, or Full Access).

Frequently asked questions

Project config or global config?

Use .cursor/mcp.json in the repo when the whole team fine-tunes on BiOS (reference an env var for the key rather than committing it). Use ~/.cursor/mcp.json for personal setup across all projects.

Does Cursor limit how many MCP tools it can load?

Cursor handles the BiOS tool set fine. If you run many MCP servers at once and hit the active-tool limit, disable servers you are not using in Settings, MCP.

Can Cursor start a training job without asking me?

By default every tool call needs your approval. Cursor has auto-run settings, but we recommend keeping approval on for tools that spend credits, like create_training_job.